Research Highlights

Elastic waves travel through bodies, encoding information along their path. Using seismometers on the Earth's surface or piezoelectric sensors in the lab, we record these waves. The record is a waveform, a source of data embedded with the wave's history. By implementing this technique on a 3/4 m lab fault, I can use the tool to image faulting processes. Currently, this is done with a 3/4 m fault of PMMA blocks sandwiching a 5 mm thick layer of quartz gouge. The unique fault setup produces a slow-slip front traveling from one fault end to the other in 1-5 s. This slow moving front shows characteristic changes in the recorded active source waveforms. The goal of this work is to resolve faulting process impacts on the waveforms so that the measurement technique may be scaled to the earth.

The Whillans Ice Plain in West Antarctica slips roughly 1/2 m once or twice daily, with little motion in-between. This cycle is resemblant of the seismic cycle common to the earthquake source physics community. My research focuses on the transition dynamics from the stuck to slipping phase. This mainshock initiation occurs via a slow-slipping front traversing the ice plain. By mimicking the spatial distribution of coupling observed on a ¾ m laboratory fault, this laboratory analog also produces slow-slipping fronts ahead of a larger mainshock. This work directly compares field data to laboratory experiments to constrain the glacial physics.

Continuum mechanics enables engineers to model complex systems, such as snow avalanches. The implementation of neural networks on domains structured as graphs offers a clever means to solve these models without the explicit definition of interpolation functions. However, it is uncertain if the trained model truly learns the underlying physics. In my Master's thesis, I evaluate if a trained graph neural network model truly captured the underlying physics inspired by the patch test introduced to the finite element community over 50 years ago. This approach shifts the evaluation of a machine learning model away from single scalar error metrics, and instead on if the model meets boundary conditions and basic physics. This approach opened the door to allow the machine learning model to be evaluated on the same metrics as conventional numerical methods, for example stability and convergence.